Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[nicalandia]

First, I can't do didly right now, as we are in a competition. The, I would have to install wine to do windows benchmarks, which I have never done before. Be patient !

Thanks, dont even worry with Wine. Its Capped at 64 threads

 

[Anhiel]

You, Sir, are very optimistic regarding Intel and very pessimistic regarding AMD, IMHO.

Numbers don't lie. Here some of my prediction numbers just for the fun. (Deviations will be CLK only; error margin 1-3%):

Ryzen 9 7940HS 8c 4/5.2GHz 35-54W
ST: 1832.5
MT:14096.3

Meteor Lake 6+8 4% IPC base ~55W; Turbo 106-157W 5.28GHz/4.95GHz/4.32GHz
based on https://www.anandtech.com/print/187...ks-review-taking-intel-s-raptor-lake-to-6-ghz
using Cinebench R23 numbers
ST: 2320.9
MT: 27081.4

Apple M3 8+2 4.25/4.12GHz/2.82GHz
ST: 2051.7
MT:17409
not reliable but since Apple likes it Geekbench5 ST:2482.3 MT:20033.9

Zen5 it's still 1 year away. Then again it's not for compete against all the above.
Zen5 8c N4 +15%IPC 8%CLK 6.1GHz/5.8GHz 90-105W TDP
ST:2543.6
MT:24185.2

Edit: Also as you can see the difference between the numbers are too large to simply overcome by a few percentages.

 

[BorisTheBlade82]

Numbers don't lie. Here some of my prediction numbers just for the fun. (Deviations will be CLK only; error margin 1-3%):

Ryzen 9 7940HS 8c 4/5.2GHz 35-54W
ST: 1832.5
MT:14096.3

Meteor Lake 6+8 4% IPC base ~55W; Turbo 106-157W 5.28GHz/4.95GHz/4.32GHz
based on https://www.anandtech.com/print/187...ks-review-taking-intel-s-raptor-lake-to-6-ghz
using Cinebench R23 numbers
ST: 2320.9
MT: 27081.4

Apple M3 8+2 4.25/4.12GHz/2.82GHz
ST: 2051.7
MT:17409
not reliable but since Apple likes it Geekbench5 ST:2482.3 MT:20033.9

Zen5 it's still 1 year away. Then again it's not for compete against all the above.
Zen5 8c N4 +15%IPC 8%CLK 6.1GHz/5.8GHz 90-105W TDP
ST:2543.6
MT:24185.2

Edit: Also as you can see the difference between the numbers are too large to simply overcome by a few percentages.

You seem to assume, that MTL will be able to clock to 6 GHz from the get go? Well, IMHO it will be IceLake-U all over again with maybe 5 GHz and an actual regression compared to their current flagship. Where you get your MT numbers is beyond me - are you assuming 55w sustained consumption? There are a lot of unknowns but being faster than a 12c 7900X @65w is a bit of a stretch as well.
And who says, that Zen5 Strix Point will top out at 8c? After PHX2 we have every reason to assume that it might be 8+X hybrid.
And from the current outlook MTL and mobile Zen5 might be presented more or less at the same time. MTL seems to have been pushed to CW52 2023 AKA 24Q1 and Strix might get presented as early as CES 24 in January.

 

[Saylick]

Numbers don't lie. Here some of my prediction numbers just for the fun. (Deviations will be CLK only; error margin 1-3%):

Ryzen 9 7940HS 8c 4/5.2GHz 35-54W
ST: 1832.5
MT:14096.3

Meteor Lake 6+8 4% IPC base ~55W; Turbo 106-157W 5.28GHz/4.95GHz/4.32GHz
based on https://www.anandtech.com/print/187...ks-review-taking-intel-s-raptor-lake-to-6-ghz
using Cinebench R23 numbers
ST: 2320.9
MT: 27081.4

Apple M3 8+2 4.25/4.12GHz/2.82GHz
ST: 2051.7
MT:17409
not reliable but since Apple likes it Geekbench5 ST:2482.3 MT:20033.9

Zen5 it's still 1 year away. Then again it's not for compete against all the above.
Zen5 8c N4 +15%IPC 8%CLK 6.1GHz/5.8GHz 90-105W TDP
ST:2543.6
MT:24185.2

Edit: Also as you can see the difference between the numbers are too large to simply overcome by a few percentages.

How did you get 2320 for CB23 ST numbers for MTL? The 12900KS gets 2350 points at 6 GHz, so even with perfect scaling and 4% higher IPC, 5.3 GHz only nets you 2150 points. Also, the 7950X which clocks to 5.7 GHz does around 2050 points, so 5.2 GHz should be around 1870. Not as big of a gap as you suggested, or am I mistaken?

 

[Geddagod]

How did you get 2320 for CB23 ST numbers for MTL? The 12900KS gets 2350 points at 6 GHz, so even with perfect scaling and 4% higher IPC, 5.3 GHz only nets you 2150 points. Also, the 7950X which clocks to 5.7 GHz does around 2050 points, so 5.2 GHz should be around 1870. Not as big of a gap as you suggested, or am I mistaken?

Should be using RPL not GLC because of L2 caches, pretty sure RWC goes to 2MB like RPC.
But ye your right, the 13700k goes to 5.3GHz, which should be around where MTL hits, and gets ~2050 points in CB23 ST, +4% is only like 2100.
Math doesn't work out.

 

[nicalandia]

WCCFTECH Just listed me as a source...

I wish they got things right for once. The AMD EPYC 9V84 is not a ES sample and the Score is for a single 96 Core CPU.

 

[Saylick]

WCCFTECH Just listed me as a source...

I wish they got things right for once. The AMD EPYC 9V84 is not a ES sample and the Score is for a single 96 Core CPU.

View attachment 79676

"Look Ma, I'm famous!"

Not surprising that WCCFTech misreports their source. It's kind of what they're known for, which is a reminder that their writers don't really care for accuracy, just volume of articles and ad revenue.

 

[Asterox]

As expected, the silicon CPU lottery or it is desirable to show such comparisons every few years.

- the worst(CPU binning)processors go to Desktop anyway

​

 

[DrMrLordX]

- the worst(CPU binning)processors go to Desktop anyway

​

The worst bins get chucked and sold to nobody. But we have no idea how many fail to even make it to desktop.

 

[Markfw]

As expected, the silicon CPU lottery or it is desirable to show such comparisons every few years.

- the worst(CPU binning)processors go to Desktop anyway

​

There is a whole thread on this. This is a report:

Discussion - Variance in CPU quality and how it affects reviews and user experience

Der8auer tested multiple R5 7600 CPUs and found up to 50% difference in efficiency while gaming. From the start everyone should know this is not related to AMD, but rather a normal occurrence in all CPUs, and the variance is obviously bigger in budget SKUs. In the past, the same youtuber did a...

forums.anandtech.com

 

[igor_kavinski]

The worst bins get chucked and sold to nobody.

That seems like such a loss. They would make great space heaters in colder climates.

Especially in laptops. Nothing like some nice radiating heat for the nads.

 

[eek2121]

I suspect Meteor Lake won't hit 6ghz. The fact it is "just" a 6+8 core part (according to rumors) suggests it won't power any flagship chips. Likely, Meteor Lake will make it's debut in low power "P" SKUs, with Raptor Lake Refresh making up the performance and budget SKUs. Doing so allows Intel to increase multicore performance of their lowest power chips without sacrificing battery life.

I could be wrong, of course. This is just speculation on my part. However, if the 6+8 part were to be in a flagship chip, it'd have to have 33% higher IPC just to make up for the loss of the two additional performance cores of the i7-13850HX, and that is to say nothing of the 4 missing efficiency cores.

By doing things in the way I suggest, Intel can avoid reaching capacity of their EUV machines while still bringing the heat to AMD.

tl;dr: I believe Meteor Lake will be mid-range only, with some defective/leaky chips making it into the low end SKUs. High end will continue to be Raptor Lake until Arrow Lake drops.

 

[nicalandia]

I suspect Meteor Lake won't hit 6ghz. The fact it is "just" a 6+8 core part (according to rumors) suggests it won't power any flagship chips. Likely, Meteor Lake will make it's debut in low power "P" SKUs, with Raptor Lake Refresh making up the performance and budget SKUs. Doing so allows Intel to increase multicore performance of their lowest power chips without sacrificing battery life.

I could be wrong, of course. This is just speculation on my part. However, if the 6+8 part were to be in a flagship chip, it'd have to have 33% higher IPC just to make up for the loss of the two additional performance cores of the i7-13850HX, and that is to say nothing of the 4 missing efficiency cores.

By doing things in the way I suggest, Intel can avoid reaching capacity of their EUV machines while still bringing the heat to AMD.

tl;dr: I believe Meteor Lake will be mid-range only, with some defective/leaky chips making it into the low end SKUs. High end will continue to be Raptor Lake until Arrow Lake drops.

Wrong thread. This is for Zen4

 

[Anhiel]

You seem to assume, that MTL will be able to clock to 6 GHz from the get go?
And who says, that Zen5 Strix Point will top out at 8c?

Due to TDP it won't hit 6GHz and that 55W was an estimate at base with room for iGPU ~20W, the value for turbo is less reliable. In any case Intel did say they can half the usage. But this could just mean lower end SKUs.
If you haven't noticed these are all mobiles not desktop comparison. And yes MTL covers mid-low range but Intel will still go all out as they do need to secure future confidence with the mess they are in now.

How did you get 2320 for CB23 ST numbers for MTL? The 12900KS gets 2350 points at 6 GHz, so even with perfect scaling and 4% higher IPC, 5.3 GHz only nets you 2150 points. Also, the 7950X which clocks to 5.7 GHz does around 2050 points, so 5.2 GHz should be around 1870. Not as big of a gap as you suggested, or am I mistaken?

I took the average of people's measured CLK which frankly wasn't always reaching 6GHz and as low as 5.58. The lowest possible value is 2220. I didn't list the 1-3% error margin for nothing.
7940HS was based on 7945HX's value ST:1903 MT:33797. In any case, 7940HS will be on N4 instead of N5 and like the Zen3+ 6000 series there's likely marginal 1% gain from the shrink/smaller path lengths without change.

Should be using RPL not GLC because of L2 caches, pretty sure RWC goes to 2MB like RPC.
But ye your right, the 13700k goes to 5.3GHz, which should be around where MTL hits, and gets ~2050 points in CB23 ST, +4% is only like 2100.
Math doesn't work out.

Yep. 4% over RPL; RPL 3% over GLC. Rumors say 20% IPC over GLC but it's likely misleading ST.

Supplemental for the iGPUs since I didn't bother and consider them a separately. It's also hard to come by wattages.

Ryzen 9 7940HS Radeon 780M 12CU 2.8GHz
8.60 TFLOPs FP32 ~15W

Meteor Lake Xe-LPG N5 128 EU 3GHz
3.072 TFLOPs FP32
Not sure where Intel gets the confidence it will match. Unless their Xe Matrix Extension can be leveraged more than I'm aware and in practice rather than general measurement.

Apple M3 assuming 12 core GPU 1.5274GHz 15W
4.72 TFLOPs FP32
M3 Pro: 20 or 23 cores GPU 7.86 or 9.05 TFLOPs FP32 20W
M3 Max: 38 or 46 cores GPU 14.95 or 18.09 TFLOPs FP32 53W

 

[ryanjagtap]

A wild phoenix point laptop appears!!!

Mainly a RDNA3 igpu benchmark, but has geekbench scores.
If you're not interested in watching the video, the scores are as follows:
Geekbench 6
Single core=2414,
multicore=9436
openCL score=30640
3DMark Firestrike -> 7367
Time Spy score -> 3190

 

[aigomorla]

Any updated dates on launch date for storm peak, and also a new WXTR board?
Did we even get updates on a new thread ripper board?
I haven't seen any Enthusiast level thread ripper board for Storm Peak yet.

 

[nicalandia]

I wonder Why Geekbench is not displaying L3$ on AMD Phoenix CPUs

 

[A///]

That seems like such a loss. They would make great space heaters in colder climates.

Especially in laptops. Nothing like some nice radiating heat for the nads.

Not needed where you live.

 

[nicalandia]

Another Custom SKU appears on Goofish. The EPYC 9R14

闲鱼宝贝详情

h5.m.goofish.com

 

[Abwx]

Some clickbait article from an usual suspect :

AMD Siena CPUs For SP6 Platform To Launch Under EPYC 8004 Branding, Up To 64 Zen 4 Cores

AMD will soon be releasing its Siena CPUs for the SP6 platform which will be branded under the EPYC 8004 family.

wccftech.com

 

[A///]

Some of it seems to be from momomo_us which has been alright in the past. I haven't seen anything from tumtum happysack in a while which is unfortunate.

 

[Abwx]

7940HS review at NBC :

https://www.notebookcheck.net/Asus-ROG-Zephyrus-G14-Review-Compact-Gaming-Laptop-with-AMD-Zen4-Phoenix-RTX-4090.711986.0.html

Test is done at CPU max power wich is 80W or so, they state that from 32W to 80W the perf delta is 40%, that s an indication that the power/frequency curve is of the form P = F^2.72.

 

[Abwx]

How to sell more powerfull GPUs and CPUs thanks to unoptimised games :

https://twitter.com/x/status/1653141414681690113

 

[Abwx]

7945HX review at Computerbase :

AMD Ryzen 9 7945HX im Test

Mit CPUs vom Typ Ryzen 7x45HX hat auch AMD die große Desktop-Plattform ins Notebook gebracht und tritt im Test gegen Intels HX-Serie an.

www.computerbase.de

 

[igor_kavinski]

Only 17% slower than a 7950X in MT!